IFIPWG7.

3Performance2013 31stInternationalSymposiumonComputerPerformance, Modeling,MeasurementsandEvaluation2013 StudentPosterAbstracts September2426,Vienna,Austria

EditorialMessagefromtheStudentPosterChair The poster proceedings contain the extended abstracts of the accepted student posters of Performance2013,heldattheUniversityofVienna,Austria,onSeptember24th26th2013.Tenposter abstracts have been received of which eight made it into the final program. Each poster abstract has been reviewed by two members of the Performance 2013 student poster tracks program committee. Sincere thanks go to the members of the program committee who provided important reviews and helpful recommendations. Further cordial thanks go to Auguste Stastny for supporting the local organization of the poster session, and Ewald Hotop for quick Web updates whenever needed. Finally, many thanks go to the poster abstract submitters for contributing to the diverse program of thestudentpostersession. September24,2013 KarinAnnaHummel,ETHZurich StudentPosterProgramChair StudentPosterProgramCommittee BernhardAger,ETHZurich GiulianoCasale,ImperialCollegeLondon AbhishekChandra,UniversityofMinnesota GeorgesdaCosta,IRIT DejanKostic,IMDEA BaochunLi,UniversityofToronto JohnC.S.Lui,TheChineseUniversityofHongKong DanielSadocMenasche,FederalUniversityofRiodeJaneiro MichelaMeo,PolitecnicodiTorino XiaZhou,DartmouthCollege

TableofContents
Poster01 HuamingWu,KatinkaWolter,andAlessandroGrazioli CloudletbasedMobileOffloadingSystems:APerformanceAnalysis Poster02 SeYoungYu,NevilBrownlee,andAniketMahanti ComparativeAnalysisofTransferProtocolsForBigData Poster03 BogdanGhitandDickEpema DynamicResourceProvisioningforConcurrentMapReduceFrameworks Poster04 SeyedSaeidMasoumzadehandHelmutHlavacs IntelligentDecisionMakinginDistributedDynamicVMConsolidationUsingFuzzyQLearning Poster05 ReneGabnerandHansPeterSchwefel EnergyefficientExecutionofMobileServicesUsingFlexibleCloudPlatforms Poster06 LjubicaPajevic,GunnarKarlssonandlafurHelgason PerformanceofEpidemicContentDistributioninOpportunisticNetworks Poster07 FlorianNiedermeier,GergLovsz,HermanndeMeer,andThomasKoppera PresentationofaSimulationToolforModelingGridStabilityintheSmartGrid Poster08 AlbertoCabrera,FranciscoAlmeida,andVincenteBlanco EML,anEnergyMeasurementLibrary

Cloudlet-based Mobile Ofoading Systems: a Performance Analysis

Huaming Wu, Katinka Wolter
Department of Mathematics and Computer Science Free University of Berlin, Berlin, Germany {huaming.wu, katinka.wolter}@fu-berlin.de

Alessandro Grazioli
Information Engineering Department Universita degli Studi di Parma, Parma, Italy alessandro.grazioli81@gmail.com

AbstractOfoading is an effective method to migrate programs from mobile devices to cloud, but it critically depends on network and cloud conditions. We suggest that the mobile device does not directly communicate with a distant cloud, but instead, with a nearby cloudlet acting as an intermediate node. Such a cloudlet-based ofoading system is modeled and analyzed with respect to the state transition process with failures. In addition, we evaluated the failure and repair time, and four types of execution time as well. The numerical results reveal that in environments characterized by higher reachability of cloudlet and cloud, longer connection time or even larger speedup factor F , this scheme benets from reduced execution time. Index Termsofoading system; cloud computing; cloudlet; network bandwidth; performance analysis

to communicate with the distant cloud, but only with the cloudlet. This model decreases latency and lowers battery consumption by using WiFi instead of broadband wireless technology.
Low-Latency high-bandwidth wireless network

mobile devices

cloudlets

distant cloud

Fig. 1. Architecture of cloudlet-based ofoading systems

I. I NTRODUCTION Ofoading a program from mobile devices to cloud is becoming increasingly attractive to reduce execution time and extend battery life. Apples Siri and iCloud [1] are two remarkable examples. However, cloud ofoading critically depends on a reliable end-to-end communication and on the availability of the cloud. In addition, it suffers from high network access latencies and low network bandwidth. Mahadev [2] proposed a vm-based cloudlet for the infrastructure setup of mobile systems. Instead, we want to investigate how effective and efcient they are and what factors inuence their performance. With this purpose, we introduce a mathematical model and analyze cloudlet-based ofoading systems with failures, considering application execution time and failure recovery time. II. S YSTEM OVERVIEW A. Problems concerning direct ofoading systems Network condition: Different network types have a large impact on communication time, cost and energy. 3G provides a near-ubiquitous coverage, but it consumes more energy than WiFi because of latencies, and is sensitive to location [3]. Cloud condition: Ofoading is difcult in locations such as the interior of a tunnel or subway, where the low network bandwidth prevents cloud applications from working properly. In addition, distant cloud dependence could lead to severe problems when service outages occur. B. Overview of cloudlet-based ofoading systems Rather than relying on a distant cloud, the resource poverty of a mobile device can be addressed by using a nearby resource-rich cloudlet via a wireless LAN. A cloudlet is a trusted, resource-rich computer which is well-connected to the internet and is available for use by nearby mobile devices [2]. As shown in Fig.1, cloudlets are dispersed and located close to mobile devices while cloud is generally far. At runtime, the app discovers a nearby cloudlet and ofoads a computationintensive program to it. The mobile device does not need

Because wireless LAN bandwidth is remarkably higher than the bandwidth provided by radio access on a mobile device [3], we choose a path connecting the mobile device to a nearby cloudlet and then to a remote cloud. As depicted in Fig.2, D is the communicated data and B is the bandwidth between the mobile device and the cloud. Likewise, B1 is the bandwidth between the mobile device and cloudlet, which generally uses gigabit internal connectivity and a high-bandwidth wireless LAN. B2 is the bandwidth between the cloudlet and cloud, which is usually based on broadband technology. Generally, B B1 and B B2 . Direct ofoading saves execution time only if Tm > Ts + D B , Tm = F Ts ; Tm and Ts are the execution times on the mobile device and cloud, respectively; and the speedup F 1 indicates how powerful the cloud is by comparison with the mobile device. Similarly, the cloudlet-based D D ofoading saves time when Tm > Ts + B +B . Therefore, the 1 2 cloudlet-based ofoading model performs better than direct 1 1 1 +B . ofoading approach when B >B 1 2
Cloudlet

B1

B2

Mobile device

B, D

Cloud

Fig. 2. Model of cloudlet-based ofoading systems

III. P ERFORMANCE ANALYSIS A. Ideal cloudlet-based ofoading systems For ideal ofoading systems, no failure occurs. The pure m (n) program execution time is Tp (n) = (1 ) Tm (n) + TF [5], here is the proportion between the sub-tasks performed by the cloud and the mobile device; 0 1, (1 ) Tm (n) m (n) and TF represent the execution time spent on the mobile device and cloud, respectively; n is the number of sub-tasks. The total execution time is TOPT (n) = Tp (n)+ Tc (n), Tc (n) = D D B1 + B2 is the total communication time.

700 Tm(n) TOPT(n) TFT(n) T(n)

B. Cloudlet-based ofoading systems with failures There are four states depicted in Fig.3. When ofoading starts, the execution state changes from SNE to SOE1 . If the remote cloud is available, it changes to SOE2 . Once the executions of all ofoaded components are successfully completed, the execution state changes back from SOE2 to SNE . However, failures may occur in all states as [5]: 1) SNE : running out of battery, abnormal shutdown. 2) SOE1 : wireless link failures, cloudlet shutdowns or becomes unreachable due to mobile devices movement. 3) SOE2 : cloud unavailable, cloud outages or becomes unreachable due to cloudlets failures. 4) SFR : nested failures may also happen.
SOE1
Offloading Execution 1
Execution time

600

500

400

300

200

100 0

10

12

14

16

18

20

Fig. 4. Execution time under different speedup F

Speedup F

1
Non-Offloading Execution Offloading Execution 2

SOE2

TFT (n) and T (n) produce a remarkably longer time than Tm (n) when F < 2, due to ofoading and failure repairs. With larger F , the cloud can save execution time. However, time spent on ofoading operation and failure handling will increase the total execution time, especially in higher cloudlet or cloud unreachability environments.
700 Tm(n) TOPT(n) 600 TFT(n) T(n)

SNE

Failure &Repair

1 2
Execution time

500

SFR

400

Fig. 3. State transitions of cloudlet-based ofoading systems with failures

300

Ofoading completes when the execution period elapses without failure. Independent failures caused by the mobile device, cloudlet and cloud are modeled as non-homogenous Poisson Processes with rates , 1 and 2 , respectively. According to Fig.3, the failure rate (t) is dened as: in state SNE and SFR + 1 in state SOE1 (t) = (1) 1 + 2 in state SOE2 The time period R is required to complete a repair in the presence of nested failures. The expectation of the failure 1 1 1 [5]. In presence repair time is E (R ) = E [eR ] of failures, the program execution time is calculated as E [TFT (n)] =

200

100 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 5. Execution time under different

As shown in Fig.5, F = 10, T (n) reduces to Tm (n) when = 0 and T (n) approaches TFT (n) when = 1. Both TFT (n) and TOPT (n) decrease along with the increase of , while T (n) rst decreases slightly and afterwards more rapidly with the increase of . TFT (n) is larger than Tm (n) when < 0.25 due to ofoading and failure recoveries. The larger is, the more execution time the cloudlet-based ofoad system saves. V. C ONCLUSION

In this short paper, we proposed an analytical model for 1 (1 1 ) E [TOE1 (n)] + 1 E [TNE/FR (n)] + cloudlet-based ofoading systems, where the mobile device 2 communicates with a nearby cloudlet instead of a remote (1 2 ) E [TOE2 (n)] + 2 E [TNE/FR (n)] (2) cloud during the entire ofoading process. In the environments characterized by high cloudlet or cloud unreachability, long disconnection time or even small speedup factor, this scheme will not benet from reduced application execution time. The analysis results provide useful guidance for the design of efcient ofoading systems. R EFERENCES
[1] D. Niu, Gearing resource-poor mobile devices with powerful clouds: architectures, challenges, and applications, IEEE Wireless Communications, vol. 3, 2013. [2] M. Satyanarayanan, P. Bahl, R. Caceres and N. Davies, The case for vm-based cloudlets in mobile computing, IEEE Pervasive Computing, vol. 8, no. 4, pp. 14-23, October 2009. [3] B. G. Chun and P. Maniatis, Dynamically partitioning applications between weak devices and clouds, in Proceedings of the 1st ACM Workshop on Mobile Cloud Computing and Services: Social Networks and Beyond (MCS), no. 7, June 2010. [4] K. Kumar and Y. H. Lu, Cloud computing for mobile users: can ofoading computation save energy? IEEE Computer, vol. 43, no. 4, pp. 51-56, April 2010. [5] S. Ou, Y. Wu, K. Yang and B. Zhou, Performance analysis of faultTolerant ofoading systems for pervasice services in mobile wireless environments, in Proceedings of the 2008 IEEE International Conference on Communications (ICC08), pp. 1856-1860, May 2008.

1 where E [TNE/FR (n)] = + E [R ] eE [TOPT (n)] 1 is the time spent in state SNE and SFR , E [TOE1 (n)] = 1 e( +1 )E [TOPT (n)] 1 in state SOE1 and +1 + E [R ] 1 E [TOE2 (n)] = 1 + e(1 +2 )E [TOPT (n)] 1 in 2 + E [R ] state SOE2 ; 1 , 2 are the probabilities of unreachability of cloudlet and cloud, respectively.

IV. N UMERICAL RESULTS AND ANALYSIS If the program ofoads, through a cloudlet, a number of sub-tasks to the cloud and the remaining ones are executed locally; we have execution time T (n) = Tm [(1 )n]+TFT (n), where Tm [(1 )n] and TFT (n) are the times spent on the mobile device and cloud. The parameters are set as follows: E (R) = 10, n = 100, = 0.9, = 103 , 1 = 104 , 2 = 105 , 1 = 0.1 and 2 = 0.2. The average execution time for each sub-task on the mobile device is 5 second and Tc = 0.3n. As shown in Fig.4, when F increases, the execution time decreases except for Tm (n), which is horizontal at 500s.

Comparative Analysis of Transfer Protocols For Big Data

[Extended Abstract]
Se-young Yu, Nevil Brownlee, and Aniket Mahanti
Department of Computer Science, University of Auckland, New Zealand

syu051@aucklanduni.ac.nz, {n.brownlee,a.mahanti}@auckland.ac.nz

ABSTRACT
Scientists working in various elds such as astronomy and biology require access to vast volumes of data (also referred to as big data) for their research. This data may need to transferred over long distances to enable collaboration among scientists around the world. It is important for them to have appropriate transfer protocols to optimize data transfer speed. This paper presents a preliminary comparative analysis of high-speed transfer protocols for big data on long fat networks. We analyze the following protocols: GridFTP, FDT, UDT and Tsunami. We perform extensive experiments to measure the ecacy of each protocol in terms of its throughput for various round-trip times, against increasing levels of congestion inducing background trac, and the impact of using jumbo frames sizes. Our results show that without much tuning, TCP-based protocols work better than UDP-based protocols.

Figure 1: 10 Gb/s Testbed Network cols, namely, GridFTP,1 FDT,2 UDT,3 and Tsunami.4 . The former two protocols are TCP-based protocols and the later two protocols are UDP-based protocols. Our goal is to understand behaviour of each of these protocols over dierent Round-Trip Times (RTT), background trac and the use of jumbo frames (i.e., frames of 9000 bytes instead of Ethernets normal 1500 bytes). Furthermore, we want to compare TCP-based and UDP-based protocols, and see which protocols perform better under stated conditions. Our preliminary analysis shows that TCP-based protocols are more ecient for large data transfers over a wide range of round trip times. When using jumbo frames, we notice increase in throughput of GridFTP on short-delay links.

1.

INTRODUCTION

Data-intensive research stations such as the Large Hadron Collider, Square Kilometre Array, and International Thermonuclear Experimental Reactor produce huge amounts of data each year. This data may need to be transferred to distant locations using high-speed networks and analyzed by researchers globally. Most research stations use 10 Gb/s links to transfer their data, however, the transfer rates they achieve using standard protocols may still be relatively low. To improve the performance of data transfer, providing only higher link capacity is not going to help much because the data transfer protocols have limitations. It is essential to choose an ecient data transfer protocol that can eciently use the available capacity, and expedite the overall transfer process. Simply relying on TCPs congestion control may result in inecient utilization of long fat pipes due to its conservative behaviour when handling packet loss events. Understanding behaviour of TCP- and UDP-based highspeed data transfer protocols will allow researchers to choose a protocol that best suits their network environment, which may dier from one research site to another. Experimental evaluation of transfer protocols in a high-speed network is lacking in previous works. In this paper, using a 10 Gb/s testbed network, we perform experiments with four well-known high-speed transfer proto-

2.

METHODOLOGY

We built a testbed network in our lab with four machines (see Figure 1) to measure behaviour of a selection of highspeed data transfer protocols. Based on results from our initial experiments, we decided to measure performance of GridFTP with TCP, GridFTP with UDT, Tsunami and FDT. Since we did not have access to a hardware trac generator, we chose to use an existing application program to generate (congestion-inducing) background trac. We used nuttcp to generate background trac. To observe their behaviour with various RTTs, we decided to use netem5 to emulate dierent delays. We tuned the test machines TCP parameters by following ESnets Linux Tuning guide6 . http://www.globus.org/toolkit/docs/latest-stable/ gridftp/ 2 http://monalisa.cern.ch/FDT/ 3 http://udt.sourceforge.net/ 4 http://tsunami-udp.sourceforge.net/ 5 http://www.linuxfoundation.org/collaborate/ workgroups/networking/netem 6 http://fasterdata.es.net/host-tuning/linux/
1

0.3 0.25 Link Utilization (%) 0.2

Tsunami GridFTP with UDT GridFTP with TCP FDT

0.15 0.1

0.05 00 50 100 Delay (ms) 150 200

With increasing UDP background trac, FDT performs very well, while GridFTP with TCP and GridFTP with UDT perform poorly. Tsunami did not work with jumbo frames and background trac, so we did not consider it further. With jumbo frames, the rate of decrease in throughput over increasing delay for FDT is reduced. As a result, FDT performs as well over long RTT links as it did for those with short RTT links. Increasing background trac did not cause much problem for FDT and GridFTP with UDT, but it made GridFTP with TCP unstable. Using jumbo frames increased the throughput of GridFTP for short RTT links, but as RTT increases, throughput of GridFTP did not improve much compared to the results with normal-size frames.

Figure 2: Throughput of the protocols using normal frames with varying RTT, with no background trac [1]

4.

CONCLUSIONS AND FUTURE WORK

3.

RESULTS

We measure throughput of each protocol when they are used to transfer data in our testbed network with varying RTT as well as varying level of background trac with varying RTT. We also measure the impact of using jumbo frames.

3.1

Impact of increasing distance

Figure 2 shows how throughput varies with RTT for the four protocols. We observe that FDT performs very well for short RTT links, while GridFTP performs better with relatively longer RTT links. Without congestion, performance of FDT and GridFTP with TCP is higher than Tsunami and UDT. The highest throughput of FDT is 2.34 Gb/s with a 1 ms RTT but it decreases rapidly after 100 ms compared to GridFTP, which performs better than FDT when link RTT is longer than 100 ms. Interestingly, the throughput of Tsunami did not decrease over increasing RTT, showing it has the most eective congestion control with increasing RTT.

We presented a preliminary comparative analysis of four well-known transfer protocols for big data, namely GridFTP, FDT, UDT, and Tsunami. Our results show that using GridFTP with TCP or FDT gives higher throughput over a 10 Gb/s network link. Even with signicant level of congestion induced by background trac, FDT was stable enough to perform data transfer. On the other hand, UDP-based protocols did not perform well in most tests and we do not recommend using Tsunami due to its instability when using jumbo frames or when faced with background trac. Our results can be used by scientists to choose the most appropriate protocol for transferring their data over long distances. There are several avenues for future work. During the experiment, we have found diculties with generating enough TCP background trac with generally accepted trac generator including Iperf, Netperf, and nuttcp. When the number of TCP ows increases, the ows interfere with each other. There are other aspects of protocols to be measured such as TCP friendliness or eect of using multiple concurrent connections. We expect to perform a further measurement for such aspects in the near future We also plan to expand our testbed over long haul links, so as to determine how best to transfer big data over long distances over production networks.

3.2

Impact of jumbo frames, background trafc, and increasing RTT

5.

REFERENCES

Figure 3 shows how throughput varies when using jumbo frames with increasing RTT for GridFTP with TCP, GridFTP with UDT and FDT.
Throughput (Gb/s)
3 2.5 2 1.5 1 0.5 9 10 7 8 100 5 6 4 3 Delay (ms) 150 1 2 UDP background 200 0 throughput (Gb/s) 50 0 0
2.8 2.6 2.4 2.2 2 1.8 1.6 1.4 1.2 1 0.8 0.6

[1] S. Yu, N. Brownlee, and A. Mahanti. Comparative performance analysis of high-speed transfer protocols for big data, Apr. 2013. submitted for publication.
Throughput (Gb/s)
2.2 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2

Throughput (Gb/s)
3 2.5 2 1.5 1 0.5 0 0

3 2.5 2 1.5 1 0.5 0 0

2.35 2.3 2.25 2.2 2.15 2.1 2.05

10 8 9 7 5 6 100 3 4 2 UDP background 150 Delay (ms) 0 1 throughput (Gb/s) 200 50

10 8 9 6 7 5 4 100 2 3 UDP background 150 Delay (ms) 0 1 200 throughput (Gb/s) 50

Figure 3: Throughput of GridFTP with TCP (left), GridFTP with UDP (middle), and FDT (right) using jumbo frames with increasing RTT and UDP background trac

Dynamic Resource Provisioning for Concurrent MapReduce Frameworks

Bogdan Ghi t
Delft University of Technology the Netherlands

Dick Epema
Delft University of Technology the Netherlands

b.i.ghit@tudelft.nl

d.h.j.epema@tudelft.nl

ABSTRACT
Running multiple instantiations of the MapReduce framework (MR-clusters) concurrently in a multicluster system or data center enables workload and data isolation, which is attractive for many organizations. We provision MR-clusters such that they receive equal levels of service by assigning each such cluster a dynamically changing weight that indicates its fair share of the resources.

generation and that consists of six clusters distributed in institutes and organizations across the Netherlands. Koala [3] is a grid resource manager developed for multicluster systems such as the DAS-4 with the goal of designing, implementing, and analyzing scheduling strategies for various application types (e.g., map-reduce, cycle scavenging jobs, workows). We have recently extended Koala with a MapReduce runner [1], which is a specialized module for scheduling and deploying MR-clusters on demand.

1.

INTRODUCTION

3.

FAIR RESOURCE ALLOCATIONS

Despite the high scalability of the MapReduce framework in large infrastructures such as multicluster systems and data centers, isolating MapReduce workloads and their data is very attractive for many organizations. In this paper, we propose a dynamic resource management approach for provisioning multiple MR-clusters in a single multicluster or data center, such that distinct MR-clusters receive equal levels of service. Towards this end, we dierentiate each MR-cluster by assigning a dynamically changing weight to it that indicates its fair share relative to all active MR-clusters. Running multiple MR-clusters concurrently within the same physical infrastructure enables four types of isolation, viz. with respect to performance, to data management, to fault tolerance, and to version [1]. Deploying each MR-cluster on a static partition of the system may lead to poor resource utilization, as some MR-clusters may be accessed more frequently than others, creating an imbalance between the levels of service they receive. To dynamically change the resource allocations at runtime, we need to understand which factors can be used to dierentiate the MR-clusters. Assuming no prior knowledge about the workloads, we propose three factors that can be used to establish the fair shares of the MR-clusters: the size of their workloads, the utilization of their allocated resources, and their service levels. As MapReduce is usually employed for data-intensive applications, one of the main issues is to limit the overhead of reorganizing the data when growing or shrinking an MRcluster when its share changes.

Our basic goal when provisioning resources to multiple MR-clusters in a single multicluster or data center is to give them equal levels of service (e.g., throughput, response time). In order to achieve this, we want to assign each MR-cluster a dynamically changing weight that indicates the share of the resources it is entitled to.

3.1

System Model

When growing or shrinking an MR-cluster, both the storage and compute layers need to be adjusted. Consequently, we distinguish three types of nodes in the system. The core nodes are fully functional nodes allocated for the MR-cluster deployment, with components in both layers of the MRcluster. The transient nodes are temporary compute nodes used to grow MR-clusters after their initial deployment. Removing transient nodes does not change the distribution of the data. The transient-core nodes are temporary fully functional nodes, also used to grow active MR-clusters. Their removal requires replicating the data they locally store.

3.2

Notion of Fairness

The MR-clusters are entitled to shares of the total data center capacity proportional to their given weights. For an MR-cluster i, the dierence between the fraction ri (t) of resources it currently has and the share of resources it should have based on its weight wi (t) at time t is dened as its temporal discrimination [2]: Di (t) = ri (t) wi (t). We dene the discrimination of MR-cluster i during a time interval [t1 , t2 ] by
t2

2.

BACKGROUND
Di (t1 , t2 ) =

In this paper we take an experimental approach to resource provisioning MR-clusters. The testbed for our experiments is the multicluster system DAS-4 1 , which is a wide-area computer system dedicated to research in parallel and distributed computing that is currently in its fourth
1

(ri (t) wi (t))dt.

t1

(1)

http://cs.vu.nl/das4

Setting t1 = ai and t2 = di with ai and di the moments of the request for the deployment and the shutdown of the MRcluster, respectively, we obtain the overall discrimination of the MR-cluster. The fairness of the system over a certain interval is indicated by the global discrimination, which is

120 100 Time [s] 80 60 40 20 0

c=20 c=30 c=40 Time [s]

140 100 80 60 40 20 0

IB 120 GbE Time [s]

1000 T J TR 750 500 250 0

20 40 60 80 Fraction of nodes removed [%]

20 40 60 80 Fraction of nodes removed [%]

10 20 30 40 Fraction of nodes removed [%]

(a)

(b)

(c)

Figure 1: The average per-node reconguration time depending on the fraction of removed nodes for (a) MRclusters with c core nodes and for (b) two networks on MR-clusters with 20 core nodes. The job execution time TJ and the reconguration time TR when removing dierent fractions of nodes from a 20-node MR-cluster running a WordCount job, after 25% of the map tasks have completed (c). dened as the variance V ar(D) across all active MR-clusters during that interval.

350 Execution Time [s] 300 250 200 150 100 50 0

3.3

Policies for Setting Weights

f=0.25 f=0.50 f=0.75 f=1.00

We dene three sets of policies for setting the weights of active MR-clusters, which are based on: the demands of the workloads submitted to them (demand-driven, e.g., in terms of the numbers of jobs or tasks in queue), the usage of the resources they are currently allocated (usage-driven), and the service they obtain in terms of one of our metrics (service-driven). In all of these policies, the weights of the MR-clusters and the global discrimination are recomputed for every interval of length T . Only when the global discrimination exceeds a threshold are the allocations of the MR-clusters actually changed according to the new weights.

WordCount Sort MR Application

Figure 2: Running WordCount (40 GB) and Sort (50 GB) on 40-node MR-clusters with dierent fractions f of core nodes. experiment, we set up MR-clusters of dierent fractions of core nodes and we run two common MapReduce applications. We nd that WordCount scales better on transient nodes than Sort (Figure 2).

3.4

Resizing Active MR-clusters

MR-clusters with positive discrimination have to shrink their resources. There are two options for doing so: nonpreemptive, with the data replication running in parallel with the workload execution, or preemptive, with the shrinking phase starting after the running tasks are killed. MRclusters with negative discrimination have to grow their resources. Based on the type of additional nodes used, there are two options: growing with transient or with transientcore nodes.

5.

CONCLUSION

Dynamic resource provisioning to multiple instantiations of the MapReduce framework deployed in single multiclusters or data centers is of both practical and theoretical interest. In this paper we propose a form of such provisioning that targets equal levels of service for the active MR-clusters.

4.

EXPERIMENTS

6.

ACKNOWLEDGMENT

In our experiments, we assess the performance of our three sets of policies for dierent values of T and with (non)preemptive shrinking and core/transient growing. As a basis, we perform a set of micro-experiments in which we assess for instance the reconguration overhead when shrinking an active MR-cluster, and the performance of running single MapReduce applications based on the number of core and transient nodes allocated to it. For the former microexperiment, we set up MR-clusters of dierent sizes with 10 GB per node replicated 3 times. We nd that the cluster size and the fraction of nodes removed have a signicant impact on the reconguration time (Figure 1a). Furthermore, increasing the network bandwidth to 20 Gb/s improves the reconguration time with only less than 20% (Figure 1b). Shrinking an active MR-cluster while running a job increases the reconguration time (Figure 1c). For the latter micro-

This work was partially funded by the Dutch national research program COMMIT.

7.

REFERENCES

[1] B. Ghit, N. Yigitbasi, and D. Epema. Resource Management for Dynamic MapReduce Clusters in Multicluster Systems. In IEEE High Performance Computing, Networking, Storage and Analysis (SCC), SC Companion, 2012. [2] D. Raz, H. Levy, and B. Avi-Itzhak. A Resource-Allocation Queueing Fairness Measure. In SIGMETRICS 2004. [3] O. Sonmez, H. Mohamed, and D. Epema. On the Benet of Processor Co-Allocation in Multicluster Grid Systems. IEEE Trans. on Parallel and Distributed Systems, 21:778789, 2010.

Intelligent Decision-Making in Distributed Dynamic VM Consolidation Using Fuzzy Q-Learning

Seyed Saeid Masoumzadeh
Research Group Entertainment Computing University of Vienna Vienna,Austria

Helmut Hlavacs
Research Group Entertainment Computing University of Vienna Vienna,Austria

masoumzadeh@gmail.com ABSTRACT
A cloud manager deals with a dynamic multi objective optimization problem. Indeed, this problem lies in the fact that there is always a tradeo between energy and performance in a virtualized data center. Therefore, a cloud manager must be equipped with a strategy to consolidate virtual machines and congure them dynamically in a way that optimizes energy-performance tradeo in an online manner. Distributed dynamic VM consolidation strategy can be an eective one to tackle this problem. The procedure of this strategy can be decomposed into four decision-making tasks.1) Host overloading detection; 2) VM selection; 3) Host underloading detection; and 4) VM placement. The dynamic optimization is achieved when each of aforementioned decisions are made optimally in an online manner. In this paper with concentration on host overloading detection and VM selection task, we propose the Fuzzy Q-Learning (FQL) as an intelligent and online machine learning approach in order to make optimal decisions towards dynamic energyperformance tradeo optimization.

helmut.hlavacs@univie.ac.at
employ a strategy to reallocate resources to virtual machines by live migration dynamically in order to maximized perserver utilization while attaining the promised non-functional qualities of the service guaranteed in the Service Level Agreements (SLA). In other word, this strategy must be able to optimize energy-performance trade-o in an online manner. Distributed dynamic VM consolidation can be an eective strategy to tackle this problem [1]. The procedure of this strategy can be decomposed into four decision-making tasks: (1) Host overloading detection, here deciding when a host must be considered as overloaded. In this situation, one or more VMs must be migrated away from this host. (2) Host underloading detection, here deciding when a host must be considered to be underloaded. In this situation, the host is ready to switch to sleep mode and all VMs must be migrated away from this host. (3) VM selection, here deciding, which VMs should be migrated away from overloaded hosts, and (4) VM placement, here deciding about which host must be selected to receive migrating VMs. Indeed, the energyperformance trade-o optimization is achieved when each of aforementioned decisions are made optimally in an online manner. Q-learning as a model free Reinforcement Learning (RL) is an eective approach to design dynamic system managements and consequently produce dynamic decision-making tasks. In principle, Q-learning can automatically learn highquality decision making tasks without an explicit model, and with little or no built-in system specic knowledge. Fuzzy Q-Learning (FQL) [2] is the fuzzy extension of Q-Learning. It is capable of handling continuity in the state space; in addition, it is powerful enough to tackle the curse of dimensionality and other ordinary RL issue rising in real life and industrial problem. In this work we employ FQL approach in order to give an intelligent behavior to the physical hosts for making optimal decision into the dynamic VM consolidation procedure towards energy-performance tradeo optimization. Several researches have been done so as to enhance the dynamic VM consolidation and several algorithms and policies have been proposed for host overloading detection and selecting VM until now. In contrast to previous studies, our proposed approach can achieve better result in improving energy-performance tradeo in cloud data center.

Categories and Subject Descriptors

D.4.7 [Operating Systems]: Organization and Design Distributed systems ; I.2.11 [Articial Intelligence]: Distributed Articial IntelligenceIntelligence Agent,Multiagent systems

General Terms
Algorithms,Management

Keywords
Dynamic VM Consolidation, Energy Ecient Cloud Manager, Fuzzy Q-Learning

1.

INTRODUCTION

Unfortunately, server consolidation enabled by virtualization introduces a new problem to the cloud environment. Since the size of Virtual Machines (VMs) inside a physical server might change due to dynamic workloads, resource underutilization or overutilization might occur, causing either inecient energy consumption or unwanted performance degradation. Consequently, the cloud managing system needs to IFIP WG 7.3 Performance 2013, September 24-26, Vienna, Austria. Copyright is held by author/owner(s).

2.

DATA CENTER AND FQL AGENTS

In our data center each physical host is associated with two FQL agents. One of them is responsible for making decision about when a host must be considered as an overloaded host and another one is responsible for making deci-

sion about which VM must be selected to migrate when a host is overloaded. All the agents inside data center by interacting whit their own corresponding host learn how to make decisions and consequently take actions according to dynamic characteristics of physical host state towards energyperformance trade-o optimization. Since the agents inside data center may experiment similar states to make decision during their lifetime; thus, we propose a cooperative learning strategy by sharing learning knowledge among agents. It can help to increase learning convergence rate and consequently make better decisions.

Figure 1: Comparision of VM slection strategies

3.

INTELLIGENT DECISION-MAKING IN HOST OVERLOADING DETECTION

Each learning agent i at each time-step t, observes the current state of its own corresponding host Sti , and gathers some information including average CPU utilization and number of VMs residing on the host as the input state Xti , where 1 i N and N is the number of hosts inside the data center. Further, the learning agent immediately chooses a threshold value as an action from the actions set A(Xt ). The threshold value is used in the host overloading detection procedure, if the CPU utilization exceeds the threshold; it means the host is overloaded and one or more VMs must be selected to migrate. One time-step later, in parts of a consequence action, the FQL agent receives a numerical reward, Rewardit+1 and nds itself in a new state St+1i . The reward is a feedback on the resulted new threshold value from the actions set. The feedback is provided by a tradeo model (we will address it in section 5) that is able to capture both energy and performance at each interval in a host. In a trial and error interaction, nally the FQL agent learns how to map the sates to the actions, so as to maximize the discounted sum of reward obtained. In other words, the FQL agent learns the best sequence of selecting thresholds, corresponding to the states observed during its lifetime.

Figure 2: Comparison of host overloading detection strategies encompass both performance degradation due to host overloading and VM migration in each host.The EC [1] is dened as a linear function of the CPU utilization of each host. ESVi = SLAVi ECi 1iN (1)

6.

EXPERIMENTAL RESULTS

4.

INTELLIGENT DECISION-MAKING IN VM SELECTION

When a host becomes overloaded, a VM selection task is called so as to remedy this situation by choosing one or more VMs for migration. Several criteria have been proposed to select VMs in the previous literatures until now .The FQL agent is able to make decision about which criterion can be eective at each state of its corresponding host for selecting VM towards energy-performance tradeo optimization. The FQL procedure is like what we discussed in the previous section; however, the denition of action set is dierent. In fact, each element of the actions set A(Xt ) is a criterion for selecting VM, such as minimum CPU utilization and maximum CPU utilization. The FQL agent in a trial and error interaction learns the best sequence of criteria, corresponding to the states observed during its lifetime.

We used CloudSim toolkit to evaluate our experiments. At the rst experiment, we have compared the result of the dynamic VM consolidation procedure when its VM selection task uses the FQL strategy with ones use minimum utilization and maximum utilization as a xed criterion. At the second experiment, we have compared the result of the dynamic VM consolidation procedure when its host overloading detection task uses the FQL strategy with the state of the art algorithms [1] that use some adaptive heuristics, based on statistical analysis of historical data for estimating CPU utilization threshold. In the comparison we have measured the ESV (Energy SLA Violation) value [1]. These experiments have been evaluated by three dierent workloads in tree dierent days. The results of these experiments have been illustrated in gure 1 and 2 respectively.Both results show the superiority of the FQL as an intelligent decision making strategy in improving energy-performance trade-o in comparison with the state of the art strategies.

7.

REFERENCES

5.

ENERGY-PERFORMANCE TRADE-OFF MODEL

To represent the energy-performance trade-o, we use the product combination of SLA Violation (SLAV) and energy consumption (EC). The SLAV [1] is a metric , which can

[1] A. Beloglazov and R. Buyya, Optimal online deterministic algorithms and adaptive heuristics for energy and performance ecient dynamic consolidation of virtual machines in cloud data centers, Concurrency and Computation: Practice and Experience, vol. 24, no. 13, pp. 13971420, 2012. [2] L. Joue, Fuzzy inference system learning by reinforcement methods, Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, vol. 28, no. 3, pp. 338355, 1998.

Energy-efcient Execution of Mobile Services using Flexible Cloud Platforms

Rene Gabner* gabner@ftw.at
*

Hans-Peter Schwefel* schwefel@ftw.at, hps@es.aau.dk

Telecommunication Research Center Vienna Donau-City-Strae 1 1220 Vienna, Austria

Aalborg University / Department of Electronic Systems Niels Jernes Vej 12 9220 Aalborg, Denmark

ABSTRACT
Two recent trends are major motivators for service component migration: the upcoming use of cloud-based services and the increasing number of mobile users accessing the Internet-based services via wireless networks. While cloudbased services target the vision of Software as a Service, where services are ubiquitously available, mobile use of such services impairs the battery life of the device caused by additional network communication between the cloud and the mobile device. This paper investigates service component migration between the mobile client - as part of the cloud and the infrastructure-based cloud as a means to save energy of the mobile device. Hereby, migration decisions are controlled by policies. A Markov modelling approach to calculate optimised policies with respect to energy consumption at the client is introduced and applied to an example scenario.

2.

SYSTEM DESCRIPTION AND EVALUATION MODEL

Mobile Device
Monitoring

Intermittent Connectivity

Cloud/ Server Platform

Run-Time
SC 1 SC 2 SC 3 SC 4

Execution Environment
SC 5 SC 6

Figure 1: Overall system architecture. The primary target of the dynamic component placement is to decrease the energy consumption at the mobile device. Therefore we adopt the approach presented in [3] to model energy consumption instead of execution time. The application is decomposed into service components, which timings are modelled using a continuous time Markov chain. All components are executed in a sequential order. We introduce a monitoring component at the client, which observes the service progress of the components, as well as the energy consumption of the device. The reconguration of the system (shift of service components from or to the client device) is triggered by the energy level observed by the monitoring component at the client and based on a pre-dened policy dening the target placement of all relevant application components. Such a reconguration process is depicted in Figure 1, where service component SC3 is just being moved from the cloud to the mobile device. We assume a factor kclient to model the slower execution of components at the client; a mean delay Drem to model additional delay for remote component calls; Df ail for delay in case of link failures (i.e. the time to recover the communication link), pf for the probability of the network to fail, and budget to set a maximum energy budget. The eort to transfer service components between client and cloud is modelled by a cost value which introduces additional delay and thus a certain energy budget, which could depend on the component size. Figure 2 shows the model of remote calls. States i and j represent local and remote execution of components. The transition from state i to state j with rate qij in the application model is in this scenario replaced by the transition structure via the 2 states in the lower part of Figure 2. Those 2 states (ij )up and (ij )down model the network states for remote calls where execution is passed to the remote component and the case of broken communication links.

1.

INTRODUCTION

Mobile cloud usage is challenged by varying wireless link conditions, as well as by limited resources at the mobile device. Therefore energy consumption of a device connected to the cloud is an important issue, both execution of a service at the client, and network communication from/to the infrastructure-based cloud have to be considered. An approach to dynamically migrate service components between the mobile device and the infrastructure-based cloud can alleviate these energy consumption problems. In doing so, the trade-o between increased energy consumption while executing components at the client, energy costs required for the component migration and reduced communication needs during execution has to be carefully balanced. While there are several frameworks e.g. [1] that allow to exibly execute applications between mobile devices and a cloud environment, the evaluation of such frameworks in simulation models [7], experimental prototypes [4], and analytic models [5] is based on heuristic choices of component migration policies. Wen et al. [6] determine optimised choices restricted to an individual application component in the setting of CPU clock frequency optimisation at the mobile device and remote invocation costs only. In contrast to that we consider real-time relocation of the whole set of components that constitute the application. The general description of a model for computation of optimised policies with respect to application execution time has been presented in previous work of the authors [2][3]. This paper extends the modelling approach to enable a new target metric for optimisation: the probability to nish within a given energy budget.

qij j i without remote interaction at server at client q'i,ij =(1-pf)qij q'i,ij =p qij q'ij ,j=(1-Drem) up down f up ijdown ijup NW down NW up q'ij =1/D fail down,ijup including network delay model for remote interaction

Figure 2: Extended service component model including network failure and execution time [2].

3.

APPLICATION TO ENERGY CONSUMPTION METRICS

Figure 3: Increased probability to nish the service in case of delayed reconguration sumed energy so far at the device. The solid curve shows a static case without reconguration while the dashed curve represents a reconguration to an optimal target conguration that was determined by the Markov model evaluation itself. So both the policy optimisation and the performance calculation are done via the model introduced in the previous section. The parameters of the scenario are shown in the title of Figure 3 and have been explained in Section 2 and 3 respectively. After the service used 3.2 J, it is the optimal time to recongure the service, which leads to an increase of the probability to nish the service within the given energy budget of 30 J by 12%.

The four dierent states in Figure 2 correspond to dierent types of activities at the mobile device and hence are associated with dierent power levels: IDLE (100mW): In state j in the example, no service computation and no messaging takes place at the client. The mobile device can operate in an energy-ecient low power mode only running the necessary background processes and hence has a low power consumption. COMPUTE (200mW): In this state, service component execution (computing) is carried out by the mobile device, while no communication takes place. Note that depending on the actually used resources on the mobile device, this single state, i in the example, could be replaced by a network of states. For the numerical example later, we however assume a single power value. TRANSCEIVE (600mW): In this state, the device is assumed to be transmitting or receiving, i.e., either due to migration of components or due to remote service component invocation. The example (ij )up in Figure 2 would actually be a transmitting state (as the control ow passes from a client component to a server component). For the numerical examples later, we use the same power level for transmitting and receiving. TRYING (800mW): In this state, (ij )down in the example, the mobile device tries to transmit, but due to some link or network problems, it is not possible. This state includes both searching for a channel, a busy physical layer, and potential retransmission attempts. To transform the time model to the energy model, we scale the transitions out of the 4 states in Figure 2 according to power values for each state. The former time instant of reconguration, called in [3], now translates to consumed energy at the mobile device in the scaled model. The deadline on application execution time in [3], now becomes a total available energy budget for this considered multicomponent application at the device. The mathematics to calculate the probability of nalising the application within the energy budget under a potential reconguration at energy consumption remains analog to the time-based model and also the computationally ecient search for an optimised target conguration remains analogously.

5.

REFERENCES

4.

NUMERICAL RESULTS

We present a result graph showing the probability for application execution within the energy budget in Figure 3. The service consists of six components with a mean execution time of one second per component, where the initial conguration is: SC1-3 at the client and SC 4-6 at a server. The x-axis is the reconguration moment, measured by con-

[1] E. Cuervo, A. Balasubramanian, D.-k. Cho, A. Wolman, S. Saroiu, R. Chandra, and P. Bahl. Maui: making smartphones last longer with code ooad. In Proceedings of the 8th international conference on Mobile systems, applications, and services, MobiSys 10, pages 4962, New York, NY, USA, 2010. ACM. [2] R. Gabner, K. A. Hummel, and H.-P. Schwefel. Modeling movable components for disruption tolerant mobile service execution. Diaz, M.; Avresky, D.; Bode, A.; Bruno, C.; Dekel, E. (Eds.) CloudComp 2009. LNCS, 34(1):231244, 2010. [3] R. Gabner, H.-P. Schwefel, K. A. Hummel, and G. Haring. Optimal model-based policies for component migration of mobile cloud services. In NCA, pages 195202, 2011. [4] T.-Y. Lin, T.-A. Lin, C.-H. Hsu, and C.-T. King. Context-aware decision engine for mobile cloud ooading. In Wireless Communications and Networking Conference Workshops (WCNCW), 2013 IEEE, pages 111116, 2013. [5] S. Ou, Y. Wu, K. Yang, and B. Zhou. Performance analysis of fault-tolerant ooading systems for pervasive services in mobile wireless environments. In Communications, 2008. ICC 08. IEEE International Conference on, pages 18561860, 2008. [6] Y. Wen, W. Zhang, and H. Luo. Energy-optimal mobile application execution: Taming resource-poor mobile devices with cloud clones. In INFOCOM, 2012 Proceedings IEEE, pages 27162720, 2012. [7] S. Yang. Manageable granularity in mobile application code ooading for energy savings. In Green Computing and Communications (GreenCom), 2012 IEEE International Conference on, pages 611614, 2012.

Performance of Epidemic Content Distribution in Opportunistic Networks

Ljubica Pajevic, Gunnar Karlsson, lafur Helgason
KTH Royal Institute of Technology Linnaeus ACCESS Center Stockholm, Sweden

{ljubica, gk, olafurr}@kth.se ABSTRACT

Epidemic spreading is a common approach to mitigate frequent link disruptions and to support content-centric information dissemination in opportunistic networks. Stochastic models, often used to characterize epidemic processes, introduce assumptions which, on one hand, make them analytically tractable, while on the other, ignore attested characteristics of human mobility. We investigate the tness of a simple stochastic model for content dissemination by comparison with experimental results obtained from real-life mobility traces. We examine four mobility datasets and consider content delivery delay as a performance metric. Our nding is that a homogeneous model is unable to capture the performance of content dissemination with respect to content delivery delays. where this distribution scheme can be used. We are interested in evaluating the performance of spreading in terms of delivery delays.

Epidemic model
We consider a homogeneous set of N mobile nodes, moving in a conned area and exchanging information through intermittent contacts. Initially, a single node carries a piece of information (a content item) and N 1 nodes are interested in obtaining the contents. We assume that the mobility of nodes is such that inter-contact times between any two nodes can be modelled by identical, independent and exponentially distributed random variables with rate . This allows analysis of the epidemic process with a stochastic model based on a continuous-time Markov chain, described in [3]. To characterize the performance, we look at two metrics: the time until the information has reached all the N nodes in the network, denoted by overall delivery time Todt , and the individual delivery time Tidt , dened as the time until an arbitrary node has obtained the contents. Their expected values, E [Todt ] and E [Tidt ] can be found from expressions given in Tab. 1. Table 1: Expected delivery times 2 Overall delivery time E [Todt ] H N N 1 1 Individual delivery time E [Tidt ] (N H 1) N 1 Hn = n i=1 1/i is the n-th harmonic number

Keywords
Epidemic modeling, content distribution, opportunistic networks, ad hoc networks

1.

INTRODUCTION

In opportunistic networks, highly dynamic network topology and intermittent connectivity make routing a challenge. A large body of algorithms for opportunistic networking employ epidemic spreading principle. Adopting the principles of epidemic modeling from the eld of mathematical biology, stochastic modeling has become a method commonly used in networking. However, for the sake of analytical tractability, models often assume identical mobility and contact patterns for all nodes in the network. Recently, studies which consider heterogeneous networks have been emerging; yet, it is debatable whether there are savings in complexity and increased understanding from using the models compared to simulations. In this poster, we present an empirical study of epidemic content spreading by using real-world mobility traces. Then, we consider an analytic model proposed in [3], and investigate if a homogeneous model can be utilized to evaluate the performance of opportunistic networks when the assumptions on network homogeneity are relaxed.

Mobility Traces
To cover various scenarios, we use four experimental datasets, diverse in time granularity, number of participants in the experiments and in the experiment duration. The datasets report direct pair-wise contacts between users moving in relatively restricted areas: a conference venue, a university, or a company building. We briey describe the contexts where the traces were collected and the acquisition methods used, and our methodology of pre-processing the traces. Infocom traces [5], obtained at Infocom 2006, report contacts between 78 experiment participants carrying iMote devices during four days. The scanning interval was 120 seconds. Humanet traces [1] describe human mobility in an oce scenario, reporting proximity traces of 52 participants during one working day. The users were carrying Bluetooth customized devices, which were scanning every 5 seconds to capture direct contacts with other devices. Supsi dataset [2] comprises of contacts in a group of 39 participants, from three institutes, located in two buildings. The participants were carrying sensor nodes with

2.

OPPORTUNISTIC CONTENT DISTRIBUTION: MODEL AND EVALUATION

The application scenario we consider is that of disseminating information by utilizing opportunistic contacts, based on user interest. Sharing local news, trac and tourist information in public areas, public announcements at massive events, or mobile advertisements are common examples

Individual delivery time

1 0.9 0.8 0.7 Infocom Humanet Supsi Milano

Overall delivery times 2.5 6 5 time (hours) 4 3 2 1 0.5 0 simulation model model
a b

Individual delivery times simulation 2 time (hours) 1.5 1 model model model
a b c

TSupsi=26 odt

hours

Tidt

Supsi

=13.8 hours

CDF

0.6 0.5 0.4 0.3 0.2 0.1 10

3

modelc

10

time (s)

Infocom

Humanet

Supsi

Milano

Infocom

Humanet

Supsi

Milano

(a) (b) (c) Figure 1: CDFs of the contents delivery times (a). Comparison of the overall (b) and individual delivery times (c). a transmission range of 5 meters and a scanning cycle of 10 milliseconds. Milano dataset [4] was obtained by 44 participants. Contacts were logged by customized radio devices, operating with a transmission range of 10 meters and a congurable scanning interval of one second. The duration of the experiments varies from a single day (Humanet ) to several weeks (Supsi ); to obtain inter-contact times, we consider only the days when a considerable number of contacts were recorded, and observe only the contact events that occurred during work hours (usually from 9:00 to 18:00). discrepancies between the analytic and simulation results are still signicant. Acknowledging the fact that contacts for many node pairs are not observable from the traces (from 9% in the Infocom to almost 70% in the Supsi trace), we propose the third method to estimate average contact rate over all node pairs, by compensating for the missing node pairs. To all node pairs which have not recorded any contacts the two nodes which havent met during the experiment we assign the same value for the average inter-contact time (for the delays in Fig. 1 that value equals the duration of the full trace), nd the average contact rate over all pairs, and compute the delivery times. These results are denoted by modelc . We observe that this method gives better estimation than the previous two. However, the inconsistency (in some scenarios the method underestimates delivery delays, while in other vastly overestimates), makes it impracticable for use in general.

Evaluation
We want to assess the capability of a homogeneous epidemic model to capture the process of epidemic content spreading in real-life scenarios. First, we simulate the epidemic process in four scenarios, and plot the CDFs of the contents delivery times in Fig. 1(a). For each of the traces, we choose a single day when the nodes were most active, seen as the number of contacts recorded during that day. The analytic model can be seen as a simple and ecient tool to estimate the performance; its simplicity stems from the fact that it requires only two input parameters: the number of nodes in the network and the node contact rate. In order to validate the analytic model, we compute the same metrics, overall and individual delivery times, given by the expressions in Tab. 1. As a rst evaluation, we assume that node interactions can be well described by the aggregate inter-contact time distributions, that is, the empirical distribution of inter-contact times estimated over all possible node pairs. From the aggregate distribution we nd the contact rate as the reciprocal of average inter-contact time. Figs. 1 (b) and (c) depict the simulation results and the expected delivery times (denoted by modela ). Clearly, this method signicantly underestimates the delivery delays, calling for more careful investigation of pair-wise node interactions. Therefore, we propose the second method to estimate the contact rate for a set of nodes. First, for every pair of nodes which reported contacts we nd the average contact rate. Then, we nd the empirical distributions of those contact rates and perform curve tting. In all four cases, log-normal distribution seems to provide the best t for the contact rate distributions. Next, we nd the average contact rates from the tted distributions and calculate the expected delivery times. Bars denoted by modelb in Fig. 1 correspond to this case. With a slight improvement from the previous method, the

3.

SUMMARY

We studied the performance of epidemic content distribution in opportunistic networks and empirically evaluated the content delivery delays by using four mobility datasets, chosen to represent a small system of users moving in a bounded area. We proposed three methods of treating the statistical data obtained from the traces, and showed that the homogeneous model is unable to accurately capture the epidemic process of the real-life scenarios. For our future work, we will aim at modeling epidemic spreading in heterogeneous systems by using other types of stochastic models.

4.

[1] J. M. Cabero, V. Molina, I. Urteaga, F. Liberal, and J. L. Martin. CRAWDAD data set Tecnalia Humanet (v. 2012-06-12), June 2012. [2] A. F orster, K. Garg, H. A. Nguyen, and S. Giordano. On context awareness and social distance in human mobility traces. In Proc.ACM, MobiOpp 12, New York, NY, USA, 2012. [3] O. Helgason, F. Legendre, V. Lenders, M. May, and G. Karlsson. Performance of opportunistic content distribution under dierent levels of cooperation. In Proc. European Wireless Conference (EW), 2010. [4] P. Meroni, S. Gaito, E. Pagani, and G. P. Rossi. CRAWDAD data set unimi/pmtr (v. 2008-12-01), Dec. 2008. [5] J. Scott, R. Gass, J. Crowcroft, P. Hui, C. Diot, and A. Chaintreau. CRAWDAD trace cambridge/haggle/imote/infocom2006, May 2009.

REFERENCES

Presentation of a Simulation Tool for Modeling Grid Stability in the Smart Grid
Florian Niedermeier, Gerg o Lov asz, Hermann de Meer and Thomas Koppera
Chair of Computer Networks and Communications Innstr. 43 94032 Passau, Germany Email: {niederme|lovasz|demeer|koppera}@m.uni-passau.de I. I NTRODUCTION providing a generic interface which allows different concrete demand response algorithms to be easily integrated. Figure 1 depicts the graphical user interface (GUI) of the simulation tool. The largest area is used by a drag&drop capable window which depicts the topology of the simulated power grid. The user may interact with this window to easily construct power networks, including power plants, transformers, households and the connections in between. The upper right window shows a graph depicting the graphs of generation and demand of power, both for real and reactive power. The middle left panel shows a histogram of the relative voltage magnitude distribution. It is used to identify the amount and duration of voltage uctuations, even in non-critical situations. The bottom right status window gives an overall status overview in form of trafc lights and signals if any simulated values (e.g., voltage) are outside the tolerable margins. The green light signals that the power grid is in a stable state whereas the orange light warns from an emerging instability in the power grid. The red light stands for a critical grid state that would lead to a blackout in the real power grid. The bottom middle and left panels give options regarding the simulation speed and view details. Topologies may be saved and opened later by using a XML-based le format [5]. The GUI of the simulation tool allows for an intuitive use of the tool. Especially the visualization of simulation runs and grid state allows to see easily how the change of certain parameters inuences the power grid. If needed, the tool can easily be extended to support a scripted input interface since the simulation tool uses an XML-based representation of power grid models. If scripted input is preferred, the GUI could still be used to visualize simulation runs and information on grid state. III. R ELATED W ORK

The future Smart Grid will support the increased feed in of decentralized renewable energy sources and an intelligent management of grid entities. However, introducing these features brings certain challenges. First, increasing the fraction of renewable energy in the power grid requires power utilities to deal with the volatile nature of solar or wind power generation. This volatility may lead to grid instabilities, especially voltage stability [1] is a key concern in areas of high solar power generation. To cope with these challenges for grid stability, voltage irregularities have to be detected as early as possible. Second, intelligent management solutions like demand response [2] are required to be highly scalable to meet the demands of large power grids. However, as the power grid is a critical infrastructure which does not tolerate long downtimes, these demand response algorithms cannot be evaluated in actual eld tests. Both challenges - a detailed grid monitoring and the possibility to evaluate new management solutions - introduce delays when implementing smart grid solutions. In case an environment does not allow for actual grid monitoring to be implemented or access to a testbed for testing smart grid algorithms does not exist, simulation may provide a feasible alternative. II. S IMULATION TOOL

In this abstract we present a work-in-progress power grid simulation tool which tackles both challenges. The tool is developed as open source software allowing the adaptation and extension of the tool. The tool allows modeling a power grid with different voltage levels via an user-friendly drag&drop interface, including fossil power generation, renewable generation and consumers [3]. The actual power grid simulation is performed using a discrete event simulation [4] in 15 minute steps. Inside the simulation the power grid is modeled as an undirected graph G = (V, E ). V denotes a set of vertices representing producers and consumers inside the power grid. It is noteworthy that a consumer can change its role to a producer at any time to account for prosumers in the actual power grid (e.g., household with high renewable energy feedin). E denotes a set of edges connecting two vertices each (E V V ). Edges represent power lines inside the power grid with certain properties such as maximum load, resistance, length, inductance, etc. One of the main applications of the tool is analyzing power grid stability. To this end it employs a GaussSeidel power ow algorithm. Another application is the simulation of demand response interactions. This is achieved by

There are several proprietary and open source simulation tools for analyzing power systems. Here, two characteristic examples of open source power analysis tools are briey described: OpenDSS1 and GridLAB-D2 . Besides several methods for frequency domain analysis OpenDSS supports the research of smart grid and topics related to energy research. However, no graphical input and only limited output is provided by the tool. Furthermore, the multitude of functionalities supported by the tool requires an intensive training before the tool can be used. GridLAB-D is a freely available open source power distribution system simulation and analysis tool. Similarly to
1 http://www.smartgrid.epri.com/SimulationTool.aspx 2 http://www.gridlabd.org/

Fig. 1.

The main GUI of the simulation tool

OpenDSS, it provides an extensive suite of analyzing tools. GridLAB-D allows also the modeling of emerging smart grid concepts such as demand response. The training material for GridLAB-D contains several hours of video courses that are necessary for getting familiar with the complex functionalities of the tool. Further simulation tools with similar characteristics are, e.g., the proprietary tools SINCAL3 and PowerFactory4 . In contrast to existing tools, our simulation tool offers only basic power analysis functionalities, but since it is open source and plug-in-based it can be easily extended by new functionality. The graphical user interface allows to start immediately with modeling without needing intensive training to understand the tool. Therefore, the tool is ideal to be used in student laboratory courses accompanying lectures. R EFERENCES
Deutsches Intstitut fr Normung-Din, DIN EN 50160 Merkmale der Spannung in ffentlichen Elektrizittsversorgungsnetzen; Deutsche Fassung EN 50160:2010 + Cor. :2010. Berlin Wien Zrich: Beuth Verlag, Februar 2011. [2] S. von Roon and T. Gobmaier, Demand response in der industrie status und potenziale in deutschland, M unchen: Forschungsstelle f ur Energiewirtschaft eV (FfE), 2010. [3] J. Glover, M. Sarma, and T. Overbye, Power System Analysis and Design. Cengage Learning, 2012.
3 http://www.energy.siemens.com/hq/en/services/power-transmissiondistribution/power-technologies-international/software-solutions/psssincal.htm 4 http://www.digsilent.de/index.php/products-powerfactory.html

T. Lechler and B. Page, Desmo-j: An object oriented discrete simulation framework in java, in Proceedings of the 11th European Simulation Symposium, 1999, pp. 4650. [5] U. Brandes, M. Eiglsperger, M. Kaufmann, J. Lerner, and C. Pich, The graphml le format, 2000.

[4]

[1]

EML, an Energy Measurement Library

Alberto Cabrera , Francisco Almeida and Vicente Blanco

Instituto Tecnol ogico y de Energ as Renovables. ITER Granadilla, Tenerife. Spain Email: acabrera@iter.es HPC Group. ETS de Ingenier a Inform atica Universidad de La Laguna, ULL La Laguna. 38270 Tenerife. Spain Email: {falmeida, Vicente.Blanco}@ull.es

I. I NTRODUCTION Over the past years, energy aware computing has become a hot topic in High Performance Computing (HPC), as the work on the eld towards the exascale computing has been progressing. To achieve exascale performance, several aspects of current computational models are being reviewed taking into account the energy consumption. In the case of the algorithmic analysis, approaches like performance proling and analytic modeling require good measurement tools to achieve the goals without having to deal with the concrete details of every hardware architecture and their respective challenges, such as how to properly measure energy. Work in the eld has been made to solve the problem of energy measurement by authors like Bellosa [1], Ryffel [2] and Ge et al. [3], but as the work was not directly available or it was a hardware solution, we started to develop a portable library called EML (Energy Measurement Library). With this implementation it is possible to abstract the user from the measuring tools, allowing to perform easiest measurement and, depending on the grade of precision and intrusion that we can have within the experiment, choose between external measurements or code instrumentation, as shown in Figure 1. Uses for this kind of library are decision making and autotuning based on energy consumption among other possibilities. To achieve so, we offer an unied interface that has a backend with implementeations for every measurement tool, which is easy to extend. We currently have for measuring energy consumption the following implementations: PDUs Intel MSR A cluster with a measurement frontend Figure 2 illustrates an example of measurement within the core using the Intel MSR Interface. It shows ten seconds of three obtained power proles that belong to an execution of various matrix multiplications using the Intel MKL dgemm routine. Power proles PP0 and PP1 belong to the processor while DRAM is the DIMM power consumption. Another example of measurement is Figure 3. In this case, measurement corresponds to an execution of High Performance Linpack, and

Fig. 2. Energy measurement from a processor using the Intel MSR RAPL interface.

Fig. 3. Energy measurement using an external metered PDU from a monitor node.

is taken directly from pdu outlets using 4 nodes. Both outputs are logged in a le that can be formatted using timestamps to synchronize energy measurement and program execution. EML allows portability for our code in different architectures, provided that the new machine is compatible with at least one of the coded solutions. II. D ETAILS The current software design is thought to minimize the effort of adding code for new measurement tools. The simplest option is to structure the library as a Factory Method design

Fig. 1. EML overview diagram. The upper part represents different hardware congurations (A processor that has Intel MSR RAPL, an external node to monitor a cluster or direct access to a PDU. The lower part shows the two different ways of adquiring data: an external program that measures energy consumption or code instrumentation. EML acts as a interconnection layer.

pattern, with every measurement tool represented as a class implementing two possible methods: instant measurement. A method that returns the instant energy consumption measured by the hardware. interval measurement. A method that returns the energy measured in an interval dened by the user. To determine the correct interval, this method is called following the classical time measurement: before and after performing the operation that requires measurement. This involves a few challenges. First, the measurement tools we have encountered have one of the two methods directly available, so the second method has to be derived from the values obtained from the available one. From instant measurement to interval measurement the operation consists on measuring on the best possible interval given by the hardware constraints and then performing the correspondent integrating operation. The opposite, from interval measurement to instant measurement, is implemented in a similar way. Measurement is done on a given interval, then the instant value in watts is equal to the amount of energy resulting from resolving the equation E (1) s P is our instant power consumption, obtained thanks to the energy E measured every s seconds. P =

Our second challenge is to unify the different information obtained by the different energy measurement tools that could generate confusion. Given the concrete case of the Intel RAPL MSR and our metered PDU, the rst obtains energy for a processor, separated in three different measures: the core itself, the RAM DIMMs and what Intel calls the uncore which, in Sandy Bridge platforms, measures the energy of the onchip graphics processor. The latter returns the energy of a whole node. For now this is resolved by tagging the measurement taken. For future work, we pretend to extend EML capabilities to facilitate the analysis of the data offering different outputs such as XML. III. ACKNOWLEDGMENTS This work was supported by the Spanish Ministry of Education and Science through the TIN2011-24598 project and the Spanish network CAPAP-H4. R EFERENCES
[1] F. Bellosa, The benets of event: driven energy accounting in powersensitive systems, in ACM SIGOPS European Workshop. ACM, 2000, pp. 3742. [2] S. Ryffel, Lea2 p: The linux energy attribution and accounting platform, Masters thesis, Swiss Federal Institute of Technology, 2009. [3] R. Ge, X. Feng, S. Song, H.-C. Chang, D. Li, and K. W. Cameron, Powerpack: Energy proling and analysis of high-performance systems and applications, IEEE Trans. Parallel Distrib. Syst., vol. 21, no. 5, pp. 658671, 2010.